Update app.py

app.py

@@ -69,6 +69,7 @@ def generator(input_ids, generation_config, max_new_tokens):
     return generation_output
     
 def loadModel():
+    global model, tokenizer
     if model is None:
         from llama_rope_scaled_monkey_patch import replace_llama_rope_with_scaled_rope
         replace_llama_rope_with_scaled_rope()